Abstract
LiDAR devices are widely used in autonomous driving scenarios and researches on 3D point cloud achieve remarkable progress over the past years. However, deep learning-based methods heavily rely on the annotation data and often face the domain generalization problem. Unlike 2D images whose domains are usually related to the texture information, the feature extracted from the 3D point cloud is affected by the distribution of the points. Due to the lack of a 3D domain adaptation benchmark, the common practice is to train the model on one benchmark (e.g, Waymo) and evaluate it on another dataset (e.g. KITTI). However, in this setting, there are two types of domain gaps, the scenarios domain, and sensors domain, making the evaluation and analysis complicated and difficult. To handle this situation, we propose LiDAR Dataset with Cross-Sensors (LiDAR-CS Dataset), which contains large-scale annotated LiDAR point cloud under 6 groups of different sensors but with same corresponding scenarios, captured from hybrid realistic LiDAR simulator. As far as we know, LiDAR-CS Dataset is the first dataset focused on the sensor (e.g., the points distribution) domain gaps for 3D object detection in real traffic. Furthermore, we evaluate and analyze the performance with several baseline detectors on the LiDAR-CS benchmark and show its applications.
Abstract (translated)
激光雷达设备在自动驾驶场景下得到广泛应用,并在3D点云研究领域取得了近年来的显著进展。然而,基于深度学习的方法在很大程度上依赖于标注数据,并常常面临领域 generalization 的问题。与2D图像通常由其纹理信息所决定的领域不同,从3D点云中提取的特征受点分布的影响。由于缺少3D领域适应基准,通常的做法是在一个基准(例如Waymo)上训练模型,并在另一个数据集(例如KITTI)上评估它。然而,在这种情况下,有两种领域差距,即场景领域和传感器领域,使得评估和分析变得复杂和困难。为了处理这种情况,我们提出了激光雷达跨传感器数据集(LiDAR-CS Dataset),它包含从不同类型的传感器的不同组中捕获的大规模标注激光雷达点云,并使用混合现实激光雷达模拟软件捕获。据我们所知,LiDAR-CS Dataset是 real-world traffic 中传感器(例如点分布)领域差距的首个专注于3D物体检测的传感器领域数据集。此外,我们还评估和分析在LiDAR-CS基准上的几个基准检测器的性能,并展示了其应用。
URL
https://arxiv.org/abs/2301.12515